What are AI "capabilities”?
An AI’s “capabilities” are the things that that AI can do — e.g., what kinds of responses GPT-4 can return when prompted.
It can be difficult to tell what an AI system is capable of. For instance, LLMs have emergent abilities that were not theorized when these models were trained. Such abilities include being responsive to prompt engineering or playing chess (kind of).
It is also difficult to know what the downstream applications and ramifications of an AI system will be, even if we know its capabilities. By way of analogy, a steam engine is primarily capable of consuming fuel and turning it into mechanical force, which has limited utility on its own. However, the steam engine has enabled many practical and conceptual breakthroughs, stretching from steam trains to the mechanization of industry and the production of electricity. In much the same way, AI models might be a crucial part of more complex systems, such as GPT-4 being used as part of AutoGPT or a model’s weights being reused as part of a custom model.